metadata.json

{
  "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
  "model_size": "1.54B parameters",
  "model_type": "instruction-tuned LLM",
  "dimensions": 1024,
  "model2vec_version": "0.7.0",
  "custom_corpus": false,
  "note": "First-ever Model2Vec distillation of instruction-tuned LLM",
  "output_dir": "models/qwen25-deposium-1024d",
  "expected_quality": "0.75-0.85+",
  "expected_speedup": "500-1000x vs full LLM",
  "final_size_mb": 65,
  "multilingual": true,
  "instruction_aware": true,
  "context_length": 32768,
  "unique_capabilities": [
    "Instruction-aware embeddings",
    "Superior semantic understanding",
    "Conversational context",
    "Code understanding",
    "Idiom recognition"
  ],
  "advantages_vs_competitors": {
    "vs_qwen3_embedding": "10x smaller, instruction-aware",
    "vs_gemma_768d": "larger base model, instruction-tuned",
    "vs_qwen3_256d": "higher quality expected, instruction-aware"
  },
  "instruction_awareness_score": 0.9198921983524119
}